AI Designs Replication-Competent Bacteriophages: A New Tool Against Drug-Resistant Infections—and a New Biosecurity Challenge

Artificial intelligence is no longer limited to modifying a few amino acids in proteins; it has now entered the stage of designing complete viral genomes. A research team used a genomic language model to generate bacteriophage sequences and produced viruses in the laboratory that could replicate and infect bacteria. This advance could provide another avenue for tackling drug-resistant bacteria, while moving biosecurity concerns from computer-generated sequences to real, functioning biological systems.

The team used bacteriophage ΦX174—which contains only 5,386 nucleotides and encodes 11 genes—as a template. The Evo model first learned sequence patterns from more than 2 million bacteriophage genomes and was then fine-tuned using 14,466 sequences from the Microviridae family. The researchers subsequently screened thousands of candidate designs and synthesized 285 complete genomes. Experimental results showed that 16 of them formed replication-competent bacteriophages in nonpathogenic Escherichia coli.

These successful designs were not simple copies of natural viruses. Each functional genome differed from its closest natural genome by 67 to 392 mutations, some of which have not been observed in known natural sequences. The host ranges of all 16 bacteriophages remained concentrated on the laboratory E. coli C strain and closely related strains, indicating that even while extensively rewriting sequences, the model could—at least in this system—retain the necessary gene configuration and host-recognition capabilities.

The test that came closest to a medical need involved first inducing bacterial resistance to natural ΦX174. The research team reported that a mixture derived from AI-designed bacteriophages could overcome the defenses of three resistant E. coli strains within one to five passages, whereas the natural reference bacteriophage was ineffective. The viruses that successfully overcame resistance carried chimeric genomes recombined from multiple designs, suggesting that the sequence diversity generated by AI could provide more starting points for bacteriophages to adapt to different bacterial receptors.

However, this remains a long way from treating superbugs. The existing evidence comes from nonpathogenic E. coli in culture dishes and has not undergone evaluation in animal models, human trials, or for manufacturing consistency. The fact that only 16 of the 285 designs succeeded also underscores that model generation does not equal biological function. ΦX174 is also one of the smallest and easiest viral genomes to synthesize. Whether this method can be extended to larger, structurally more complex, or clinically relevant bacteriophages remains unproven.

The researchers deliberately excluded viral sequences that infect humans, animals, and plants, and restricted the experiments to nonpathogenic hosts and controlled environments. However, these measures primarily constrain the current model and experiments and cannot replace comprehensive governance. Biosecurity experts advocate a multilayered defense, including restricting access to models and data, reviewing research objectives, screening high-risk sequences during DNA synthesis, and maintaining laboratory biosafety and security measures.

The practical significance of this work, therefore, is not that AI can already arbitrarily create cancer-fighting viruses or novel pathogens, but that it clearly demonstrates for the first time that generative models can propose complete viral genomes, a small number of which are genuinely functional after synthesis and screening. Its future value will depend on whether this design capability can be directed toward specific bacteria and verifiable therapeutic needs; its risks will depend on whether safety systems can establish defenses before more powerful models and cheaper DNA synthesis technologies become widespread.